🤖 AI Brief: Superintelligence coming "this decade", US military trials 5 LLMs, and ChatGPT’s traffic decline. Plus: 1B token

Today is July 10, 2023.

Welcome to another week of fascinating AI news. Top of mind for me: will OpenAI’s prediction that superintelligence arrives “this decade” actually come true? If so — we’re nowhere near ready as a society, and OpenAI seems aware of that too.

Also: I’ve had so many people inquire about other ways to work together in the past few weeks that I’ve included a new section below. Thanks to everyone who’s emailed in with questions and feedback!

In this issue:

• 🧨 OpenAI says "superintelligence" will arrive "this decade," so they're creating the Superalignment team

• 🪖 US military now trialing 5 LLMs trained on classified data

• 💔 ChatGPT is losing users for the first time

• 🔎 Sarah Silverman sues ChatGPT, and other news scoops

• 🧪 The latest science experiments, including a 1B token Transformer architecture

Did someone forward this to you? Subscribe here!

🧨 OpenAI says "superintelligence" will arrive "this decade," so they're creating the Superalignment team

OpenAI says superintelligence (which is more capable than AGI, in their view) could arrive "this decade," and it could be "very dangerous." As a result, they're forming a new Superalignment team led by two of their most senior researchers and dedicating 20% of their compute resources to this effort.

Why this matters:

• "Superintelligence will be the most impactful technology humanity has ever invented," but human society currently doesn't have solutions for steering or controlling superintelligent AI, says OpenAI.

• Current alignment techniques don't scale to superintelligence because humans can't reliably supervise AI systems smarter than them.

How can superintelligence alignment be solved? OpenAI has an answer:

• An automated alignment researcher (an AI bot) is the solution. This means an AI system is helping align AI: in OpenAI's view, the scalability here enables robust oversight and automated identification and solving of problematic behavior.

OpenAI wants to solve this in the next four years, given they anticipate superintelligence could arrive "this decade." And they seem generally optimistic: "superintelligence alignment is fundamentally a machine learning problem, and we think great machine learning experts—even if they’re not already working on alignment—will be critical to solving it."

🪖 US military now trialing 5 LLMs trained on classified data, wants AI to empower military planning

The US military has always been interested in AI, and now they’ve revealed they’re moving quickly on testing generative AI technologies. Bloomberg reports (note: paywalled article) that the US military is currently trialing 5 separate LLMs, all trained on classified military data, in an exercise lasting most of July.

Why this matters:

• Long-term, the US wants AI to empower military planning, sensor analysis, and firepower decisions. So think of this as just a first step in their broader goals for AI over the next decade.

• There's a tremendous amount of proprietary data for LLMs to digest: information retrieval and analysis is a huge challenge for militaries -- going from boolean searching to natural language queries is already a huge step up.

What are they testing? Details are scarce, but here's what we do know:

• ScaleAI's Donovan platform is one of them. Donovan is defense-focused AI platform and ScaleAI divulged in May that the XVIII Airborne Corps would trial their LLM.

• The four other LLMs are unknown, but expect all the typical players, including OpenAI. Microsoft has a $10B Azure contract with DoD already in place.

• LLMs are evaluated for military response planning in this trial phase: they'll be asked to help plan a military response for escalating global crisis that starts small and then shifts into the Indo-Pacific region.

Early results show military plans can be completed in "10 minutes" for something that would take hours to days, a colonel has revealed.

Cue the countdown clock til SkyNet arrives.

💔 ChatGPT is losing users for the first time. Here’s what could be happening.

The Washington Post (note: paywalled article) reports that traffic data from SimilarWeb is showing the first traffic decrease for ChatGPT since its launch last November. We heavily disagree with their headline that this is “shaking faith in AI revolution.”

Remember: ChatGPT took just two months to get to 100 million monthly users (a record now smashed by Meta’s Threads app), making it the fastest growing consumer app ever at time of launch.

Let’s unpack what could be going on:

• Could interest in chatbots be waning? Our honeymoon period with their capabilities could be over as dozens of popular chatbots proliferate and we become more attuned to their weaknesses and limitations.

• At the same time, users are complaining that ChatGPT’s capabilities are getting worse. This support thread on OpenAI is one of the many examples of users sharing increased frustration that ChatGPT no longer generates quality outputs. Reddit is also filled with similar complaints.

Or, the reasons behind the decline could be innocuous as well:

• The school year is over, and usage of ChatGPT could have seen a seasonable decrease due to reduced student usage.

• And it feels like every tool these days now has generative AI integrated in it – so going directly to ChatGPT’s site is simply less needed.

🔎 Quick Scoops

All ChatGPT Plus users now have access to OpenAI’s Code Interpreter, giving everyone a data analyst copilot. (VentureBeat)

A $3B class action lawsuit was filed against OpenAI and Microsoft, alleging OpenAI secretly “scraped 300 billion words from the internet” without registering as a data broker or obtaining consent. (Vice)

Sarah Silverman sues ChatGPT and Meta, alleging copyright infringement. Her works were used as part of copyright-infringing data sets that the LLMs digested, she says. (The Verge)

Google makes a big push into medical AI. There’s big money here – but many experts think we’re not ready yet. (Wall Street Journal - note: paywalled article)

GPT-4 is actually a mix of 8 different 220-billion-parameter models, leaks reveal. Previous rumors had suggested it was a 1T parameter model, but those are now debunked. (Algorithmic Bridge - note: paywalled article)

🧪 Science Experiments

Hugging Face launches text-to-video resource page

• As open-source models proliferate, Hugging Face has assembled 33 leading open-source models and 7 datasets, along with instructions, that make it easier for anyone curious about text-to-video models to get started. Check it out here.

LongNet architecture: scaling Transformers to 1,000,000,000 tokens

• Scaling sequence length is critically important for language models, but performance issues typically arise as model size grows. This team proposes a new Transformer variant that doesn’t sacrifice performance on shorter sequences while enabling (theoretically) up to 1 billion tokens.

• arXiv page here

As open-source LLMs proliferate, it’s easy to spread a “poisoned” LLM online

• This article from Mithril Security highlights how they hid a modified version of a popular open-source LLM on HuggingFace. This “poisoned” LLM was surginally modified to encode false information. Read more here.

😀 A reader’s commentary

👋 How I can help

As this newsletter nears 15,000 subscribers (!), I’ve had the amazing chance to hear from many of you on what you want beyond just a weekly roundup.

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, Tome, Twelve Labs and more to help them make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at [email protected] — we’ll be opening up the first set of slots soon.